Natural Language Processing
Mechanistic Interpretability and Controllable AI
We study the internal mechanisms of large language models to understand how knowledge is represented, retrieved, and utilized during inference. Our research focuses on concept discovery, circuit analysis, inference-time steering, faithful reasoning, and model editing, with the goal of developing interpretable, controllable, and trustworthy AI systems.
-
Decoding Internal Concept Representations and Building Concept Dictionaries for AI
-
Circuit Analysis of Reasoning and Knowledge Retrieval Mechanisms
-
Model Steering Through Inference-Time Interventions
-
Evaluating the Faithfulness of Chain-of-Thought Reasoning and Analyzing Rationalization
-
Knowledge Updating and Selective Forgetting Through Model Editing
